Executing a Collective Operation Algorithm in a Parallel Computer

1. A method of executing a collective operation algorithm in a parallel computer, the parallel computer comprising a plurality of compute nodes, the compute nodes coupled for data communications by one or more data communications networks, the compute nodes organized in an operational group for collective operations, the method comprising: determining, by a compute node of an operational group, a required number of participants for execution of a collective operation algorithm; determining, by the compute node, a number of contributing nodes having data to participate in the algorithm; calculating, by the compute node, a number of ghost nodes to participate in the algorithm, wherein the number of ghost nodes being the required number of participants minus the number of contributing nodes having data to participate; selecting, by the compute node, the collective operation algorithm based on the calculated number of ghost nodes; selecting from a plurality of ghost nodes, by the compute node, the calculated number of ghost nodes for participation in the execution of the algorithm, wherein selecting the calculated number of ghost nodes comprises: determining, by the compute node, based on properties of the selected collective operation algorithm and rank of contributing nodes, to involve a subset of ghost nodes from the plurality of ghost nodes corresponding to one or more of the contributing nodes; and executing the algorithm with both the selected ghost nodes and the one or more contributing nodes.

2. The method of claim 1 wherein each selected ghost node is configured to participate in the algorithm with an identity operation; and wherein during execution of the algorithm, each selected ghost node participates with the identity operation.

3. The method of claim 1 further comprising mapping, by the compute node, a selected ghost node associated with the compute node to physical hardware for execution of an identity operation.

4. The method of claim 1 wherein each compute node of the distributed processing system creates during startup of the compute node, a corresponding ghost node.